float32 should be fine for mnist loss and accuracy metrics and float64 is not available on TPUs.
Attach a file by drag & drop or click to upload
